NVIDIA releases Nemotron 3.5 Lightning: 30B LatentMoE, 3B active, 1M context

huggingface · x · 2026-08-11

NVIDIA unveiled Nemotron 3.5 Lightning for more efficient agents. It features a 30B LatentMoE with 3B active parameters, supports NVFP4 and BF16, up to 1M context, MTP, DFlash, DSpark, tool use, and multilingual. Integrated with NousResearch's Hermes Agent and runs via Microsoft Foundry.

Related event: NVIDIA Open-Sources Nemotron Model and Switchyard Router(27 posts)→

Original post →

More from Models

Models channel →